In order to build a portfolio selection model for a robo-advisor, which can be used on ETFs of mainland China and get the efficient frontier, a number of models based on the mean-variance model are studied and analyzed experimentally, the results show that the hybrid model using Hopfield neural network and genetic algorithm can output efficient frontier better than others. Based on this, exponentially weighted moving average/covariance are applied to adjust the model's inputs, that is, the mean and covariance of assets's return rate. Experiments were conducted using the collected transaction data of ETFs, the results show that after the adjustment the model can know future performance of portfolios better based on long-term historical transaction data.